Differences in HDR

[aths]

what's wrong with FP10? seems like some of you guys are downplaying FP10 now. sure it's not as great as FP16 but as long as it looks good enough right? just like nao32?

In some cases, FP10 can be inferiour to FX8.

I consider 8 bit fixpoint low dynamic range, the double bit width (16 bit words) as medium dynamic range and 32 bit fixpoint as high dynamic range. With denorm support, FP16 already delivers more dynamic than 32 bit fixpoint. (While FP16 without denorm support has a smaaller dynamic range.)

 

[Fafalada]

You could blend in a shader but it would pretty much mean copying the back-buffer to a temporary texture for each polygon, and then using that as a source texture.

But that's only true if your architecture separates render buffer and texture memory (ala Xenos) and even then it's possible to do some sorting heuristics for the alpha geometry to render in non-overlapping batches, reducing the number of required resolves to texture memory to much less then one per primitive.
Of course you'd want to do this on unified memory solutions to avoid cache flushes also (unless it's a Sony GPU ).

All of which could be avoided if we did have read access to renderbuffers in PS

 

[Lycan]

"Of course you'd want to do this on unified memory solutions to avoid cache flushes also (unless it's a Sony GPU )."

Could it be...A HINT ?

One question to those who are "in touch" with PS3 hardware on everyday basis. I know there are NDA's you have to abide by, and I'am not interested in the heav-tech stuff; but, on every forum, discussions are blossoming as how the RSX might end up inferior to G70...Huh ?

 

[MrWibble]

But that's only true if your architecture separates render buffer and texture memory (ala Xenos) and even then it's possible to do some sorting heuristics for the alpha geometry to render in non-overlapping batches, reducing the number of required resolves to texture memory to much less then one per primitive.
Of course you'd want to do this on unified memory solutions to avoid cache flushes also (unless it's a Sony GPU ).

All of which could be avoided if we did have read access to renderbuffers in PS

It's annoying, but the depth of the pipeline and the extend to which reads and writes get out of order and/or combined at the back end mean that performance would take a nose-dive if direct render-target access was possible in the shader.

Even a fairly dumb architecture like PS2's GS wasn't entirely happy with the idea of rendering to the same thing you're reading from (though if you were extremely careful, it did actually work).

Your sorting could work, but you'd probably have to bucket that sort on whatever tile-size and alignment/overlap is used by the ROP stage, so you'd probably get multiple draw passes even for a relatively low over-draw operations.

 

[MrWibble]

One question to those who are "in touch" with PS3 hardware on everyday basis. I know there are NDA's you have to abide by, and I'am not interested in the heav-tech stuff; but, on every forum, discussions are blossoming as how the RSX might end up inferior to G70...Huh ?

No real comment to make on that, IMO it's just another flavour of "ABC will be better than XYZ" argument. Sometimes people just hear what they want to hear...

 

[Fafalada]

It's annoying, but the depth of the pipeline and the extend to which reads and writes get out of order and/or combined at the back end mean that performance would take a nose-dive if direct render-target access was possible in the shader.

Well I 'was' implying that shader access would involve removal of the fixed hw shenanigans at the ROP side. But yea - I'm aware of the problems with allowing this.

Even a fairly dumb architecture like PS2's GS wasn't entirely happy with the idea of rendering to the same thing you're reading from (though if you were extremely careful, it did actually work).

Well it was just a matter of flushing page buffer per primitive like you suggested earlier, but on GS the performance hit for doing so was insignificant, where as most on most modern GPUs it would completely kill performance.
And on GS if you were VRam constrained this allowed you to for instance use alpha for stencil buffer - I know a certain famous board member who used this in a game, I guess he was one of the nasty people that wanted to break backward compatibility in PS3
Don't look at me - I didn't do it - I was more fond of doing freaky things to DMA myself.

Could it be...A HINT ?

Just to avoid any confusion here - I wasn't hinting anything - I said "Sony" not "Nvidia" GPU.